[PATCH] selftests/powerpc: Add a test of spectre_v2 mitigations
Michael Ellerman
mpe at ellerman.id.au
Mon May 20 20:55:20 AEST 2019
This test uses the PMU to count branch prediction hits/misses for a
known loop, and compare the result to the reported spectre v2
mitigation.
This gives us a way of sanity checking that the reported mitigation is
actually in effect.
Sample output for some cases, eg:
Power9:
sysfs reports: 'Vulnerable'
PM_BR_PRED_CCACHE: result 368 running/enabled 5792777124
PM_BR_MPRED_CCACHE: result 319 running/enabled 5792775546
PM_BR_PRED_PCACHE: result 2147483281 running/enabled 5792773128
PM_BR_MPRED_PCACHE: result 213604201 running/enabled 5792771640
Miss percent 9 %
OK - Measured branch prediction rates match reported spectre v2 mitigation.
sysfs reports: 'Mitigation: Indirect branch serialisation (kernel only)'
PM_BR_PRED_CCACHE: result 895 running/enabled 5780320920
PM_BR_MPRED_CCACHE: result 822 running/enabled 5780312414
PM_BR_PRED_PCACHE: result 2147482754 running/enabled 5780308836
PM_BR_MPRED_PCACHE: result 213639731 running/enabled 5780307912
Miss percent 9 %
OK - Measured branch prediction rates match reported spectre v2 mitigation.
sysfs reports: 'Mitigation: Indirect branch cache disabled'
PM_BR_PRED_CCACHE: result 2147483649 running/enabled 20540186160
PM_BR_MPRED_CCACHE: result 2147483649 running/enabled 20540180056
PM_BR_PRED_PCACHE: result 0 running/enabled 20540176090
PM_BR_MPRED_PCACHE: result 0 running/enabled 20540174182
Miss percent 100 %
OK - Measured branch prediction rates match reported spectre v2 mitigation.
Power8:
sysfs reports: 'Vulnerable'
PM_BR_PRED_CCACHE: result 2147483649 running/enabled 3505888142
PM_BR_MPRED_CCACHE: result 9 running/enabled 3505882788
Miss percent 0 %
OK - Measured branch prediction rates match reported spectre v2 mitigation.
sysfs reports: 'Mitigation: Indirect branch cache disabled'
PM_BR_PRED_CCACHE: result 2147483649 running/enabled 16931421988
PM_BR_MPRED_CCACHE: result 2147483649 running/enabled 16931416478
Miss percent 100 %
OK - Measured branch prediction rates match reported spectre v2 mitigation.
success: spectre_v2
Signed-off-by: Michael Ellerman <mpe at ellerman.id.au>
---
.../testing/selftests/powerpc/include/utils.h | 1 +
.../selftests/powerpc/security/Makefile | 3 +-
.../selftests/powerpc/security/branch_loops.S | 82 +++++++
.../selftests/powerpc/security/spectre_v2.c | 218 ++++++++++++++++++
tools/testing/selftests/powerpc/utils.c | 20 ++
5 files changed, 323 insertions(+), 1 deletion(-)
create mode 100644 tools/testing/selftests/powerpc/security/branch_loops.S
create mode 100644 tools/testing/selftests/powerpc/security/spectre_v2.c
diff --git a/tools/testing/selftests/powerpc/include/utils.h b/tools/testing/selftests/powerpc/include/utils.h
index 7636bf45d5d5..be493b4823f7 100644
--- a/tools/testing/selftests/powerpc/include/utils.h
+++ b/tools/testing/selftests/powerpc/include/utils.h
@@ -34,6 +34,7 @@ int pick_online_cpu(void);
int read_debugfs_file(char *debugfs_file, int *result);
int write_debugfs_file(char *debugfs_file, int result);
+int read_sysfs_file(char *debugfs_file, char *result, size_t result_size);
void set_dscr(unsigned long val);
int perf_event_open_counter(unsigned int type,
unsigned long config, int group_fd);
diff --git a/tools/testing/selftests/powerpc/security/Makefile b/tools/testing/selftests/powerpc/security/Makefile
index 85861c46b445..d68e6031695c 100644
--- a/tools/testing/selftests/powerpc/security/Makefile
+++ b/tools/testing/selftests/powerpc/security/Makefile
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: GPL-2.0+
-TEST_GEN_PROGS := rfi_flush
+TEST_GEN_PROGS := rfi_flush spectre_v2
top_srcdir = ../../../../..
CFLAGS += -I../../../../../usr/include
@@ -8,3 +8,4 @@ CFLAGS += -I../../../../../usr/include
include ../../lib.mk
$(TEST_GEN_PROGS): ../harness.c ../utils.c
+$(OUTPUT)/spectre_v2: ../pmu/event.c branch_loops.S
diff --git a/tools/testing/selftests/powerpc/security/branch_loops.S b/tools/testing/selftests/powerpc/security/branch_loops.S
new file mode 100644
index 000000000000..22e9204e3421
--- /dev/null
+++ b/tools/testing/selftests/powerpc/security/branch_loops.S
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: GPL-2.0+
+
+/*
+ * Copyright 2019, Michael Ellerman, IBM Corp.
+ */
+
+#include <ppc-asm.h>
+
+ .data
+
+jump_table:
+ .long 0x0
+ .long (.Lstate_1 - .Lstate_0)
+ .long (.Lstate_2 - .Lstate_0)
+ .long (.Lstate_3 - .Lstate_0)
+ .long (.Lstate_4 - .Lstate_0)
+ .long (.Lstate_5 - .Lstate_0)
+ .long (.Lstate_6 - .Lstate_0)
+ .long (.Lstate_7 - .Lstate_0)
+
+ .text
+
+#define ITER_SHIFT 31
+
+.macro state number
+ .balign 32
+.Lstate_\number:
+ .if \number==7
+ li r3, 0
+ .else
+ li r3, \number+1
+ .endif
+ b .Lloop
+.endm
+
+FUNC_START(pattern_cache_loop)
+ li r3, 0
+ li r4, 1
+ sldi r4, r4, ITER_SHIFT
+
+.Lloop: cmpdi r4, 0
+ beqlr
+
+ addi r4, r4, -1
+
+ ld r6, jump_table at got(%r2)
+ sldi r5, r3, 2
+ lwax r6, r5, r6
+ ld r7, .Lstate_0 at got(%r2)
+ add r6, r6, r7
+ mtctr r6
+ bctr
+
+ state 0
+ state 1
+ state 2
+ state 3
+ state 4
+ state 5
+ state 6
+ state 7
+
+FUNC_END(pattern_cache_loop)
+
+
+FUNC_START(indirect_branch_loop)
+ li r3, 1
+ sldi r3, r3, ITER_SHIFT
+
+1: cmpdi r3, 0
+ beqlr
+
+ addi r3, r3, -1
+
+ ld r4, 2f at got(%r2)
+ mtctr r4
+ bctr
+
+ .balign 32
+2: b 1b
+
+FUNC_END(indirect_branch_loop)
diff --git a/tools/testing/selftests/powerpc/security/spectre_v2.c b/tools/testing/selftests/powerpc/security/spectre_v2.c
new file mode 100644
index 000000000000..8c6b982af2a8
--- /dev/null
+++ b/tools/testing/selftests/powerpc/security/spectre_v2.c
@@ -0,0 +1,218 @@
+// SPDX-License-Identifier: GPL-2.0+
+
+/*
+ * Copyright 2018-2019 IBM Corporation.
+ */
+
+#define __SANE_USERSPACE_TYPES__
+
+#include <sys/types.h>
+#include <stdint.h>
+#include <malloc.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <sys/prctl.h>
+#include "utils.h"
+
+#include "../pmu/event.h"
+
+
+extern void pattern_cache_loop(void);
+extern void indirect_branch_loop(void);
+
+static int do_count_loop(struct event *events, bool is_p9, s64 *miss_percent)
+{
+ u64 pred, mpred;
+
+ prctl(PR_TASK_PERF_EVENTS_ENABLE);
+
+ if (is_p9)
+ pattern_cache_loop();
+ else
+ indirect_branch_loop();
+
+ prctl(PR_TASK_PERF_EVENTS_DISABLE);
+
+ event_read(&events[0]);
+ event_read(&events[1]);
+
+ // We could scale all the events by running/enabled but we're lazy
+ // As long as the PMU is uncontended they should all run
+ FAIL_IF(events[0].result.running != events[0].result.enabled);
+ FAIL_IF(events[1].result.running != events[1].result.enabled);
+
+ pred = events[0].result.value;
+ mpred = events[1].result.value;
+
+ if (is_p9) {
+ event_read(&events[2]);
+ event_read(&events[3]);
+ FAIL_IF(events[2].result.running != events[2].result.enabled);
+ FAIL_IF(events[3].result.running != events[3].result.enabled);
+
+ pred += events[2].result.value;
+ mpred += events[3].result.value;
+ }
+
+ *miss_percent = 100 * mpred / pred;
+
+ return 0;
+}
+
+static void setup_event(struct event *e, u64 config, char *name)
+{
+ event_init_named(e, config, name);
+
+ e->attr.disabled = 1;
+ e->attr.exclude_kernel = 1;
+ e->attr.exclude_hv = 1;
+ e->attr.exclude_idle = 1;
+}
+
+enum spectre_v2_state {
+ VULNERABLE = 0,
+ UNKNOWN = 1, // Works with FAIL_IF()
+ NOT_AFFECTED,
+ BRANCH_SERIALISATION,
+ COUNT_CACHE_DISABLED,
+ COUNT_CACHE_FLUSH_SW,
+ COUNT_CACHE_FLUSH_HW,
+ BTB_FLUSH,
+};
+
+static enum spectre_v2_state get_sysfs_state(void)
+{
+ enum spectre_v2_state state = UNKNOWN;
+ char buf[256];
+ int len;
+
+ memset(buf, 0, sizeof(buf));
+ FAIL_IF(read_sysfs_file("devices/system/cpu/vulnerabilities/spectre_v2", buf, sizeof(buf)));
+
+ // Make sure it's NULL terminated
+ buf[sizeof(buf) - 1] = '\0';
+
+ // Trim the trailing newline
+ len = strlen(buf);
+ FAIL_IF(len < 1);
+ buf[len - 1] = '\0';
+
+ printf("sysfs reports: '%s'\n", buf);
+
+ // Order matters
+ if (strstr(buf, "Vulnerable"))
+ state = VULNERABLE;
+ else if (strstr(buf, "Not affected"))
+ state = NOT_AFFECTED;
+ else if (strstr(buf, "Indirect branch serialisation (kernel only)"))
+ state = BRANCH_SERIALISATION;
+ else if (strstr(buf, "Indirect branch cache disabled"))
+ state = COUNT_CACHE_DISABLED;
+ else if (strstr(buf, "Software count cache flush (hardware accelerated)"))
+ state = COUNT_CACHE_FLUSH_HW;
+ else if (strstr(buf, "Software count cache flush"))
+ state = COUNT_CACHE_FLUSH_SW;
+ else if (strstr(buf, "Branch predictor state flush"))
+ state = BTB_FLUSH;
+
+ return state;
+}
+
+#define PM_BR_PRED_CCACHE 0x040a4 // P8 + P9
+#define PM_BR_MPRED_CCACHE 0x040ac // P8 + P9
+#define PM_BR_PRED_PCACHE 0x048a0 // P9 only
+#define PM_BR_MPRED_PCACHE 0x048b0 // P9 only
+
+#define SPRN_PVR 287
+
+int spectre_v2_test(void)
+{
+ enum spectre_v2_state state;
+ struct event events[4];
+ s64 miss_percent;
+ bool is_p9;
+
+ state = get_sysfs_state();
+ if (state == UNKNOWN) {
+ printf("Error: couldn't determine spectre_v2 mitigation state?\n");
+ return -1;
+ }
+
+ memset(events, 0, sizeof(events));
+
+ setup_event(&events[0], PM_BR_PRED_CCACHE, "PM_BR_PRED_CCACHE");
+ setup_event(&events[1], PM_BR_MPRED_CCACHE, "PM_BR_MPRED_CCACHE");
+ FAIL_IF(event_open(&events[0]));
+ FAIL_IF(event_open_with_group(&events[1], events[0].fd) == -1);
+
+ is_p9 = ((mfspr(SPRN_PVR) >> 16) & 0xFFFF) == 0x4e;
+
+ if (is_p9) {
+ // Count pattern cache too
+ setup_event(&events[2], PM_BR_PRED_PCACHE, "PM_BR_PRED_PCACHE");
+ setup_event(&events[3], PM_BR_MPRED_PCACHE, "PM_BR_MPRED_PCACHE");
+
+ FAIL_IF(event_open_with_group(&events[2], events[0].fd) == -1);
+ FAIL_IF(event_open_with_group(&events[3], events[0].fd) == -1);
+ }
+
+ FAIL_IF(do_count_loop(events, is_p9, &miss_percent));
+
+ event_report_justified(&events[0], 18, 10);
+ event_report_justified(&events[1], 18, 10);
+ event_close(&events[0]);
+ event_close(&events[1]);
+
+ if (is_p9) {
+ event_report_justified(&events[2], 18, 10);
+ event_report_justified(&events[3], 18, 10);
+ event_close(&events[2]);
+ event_close(&events[3]);
+ }
+
+ printf("Miss percent %lld %%\n", miss_percent);
+
+ switch (state) {
+ case VULNERABLE:
+ case NOT_AFFECTED:
+ case COUNT_CACHE_FLUSH_SW:
+ case COUNT_CACHE_FLUSH_HW:
+ // These should all not affect userspace branch prediction
+ if (miss_percent > 15) {
+ printf("Branch misses > 15%% unexpected in this configuration!\n");
+ printf("Possible mis-match between reported & actual mitigation\n");
+ return 1;
+ }
+ break;
+ case BRANCH_SERIALISATION:
+ // This seems to affect userspace branch prediction a bit?
+ if (miss_percent > 25) {
+ printf("Branch misses > 25%% unexpected in this configuration!\n");
+ printf("Possible mis-match between reported & actual mitigation\n");
+ return 1;
+ }
+ break;
+ case COUNT_CACHE_DISABLED:
+ if (miss_percent < 95) {
+ printf("Branch misses < 20%% unexpected in this configuration!\n");
+ printf("Possible mis-match between reported & actual mitigation\n");
+ return 1;
+ }
+ break;
+ case UNKNOWN:
+ case BTB_FLUSH:
+ printf("Not sure!\n");
+ return 1;
+ }
+
+ printf("OK - Measured branch prediction rates match reported spectre v2 mitigation.\n");
+
+ return 0;
+}
+
+int main(int argc, char *argv[])
+{
+ return test_harness(spectre_v2_test, "spectre_v2");
+}
diff --git a/tools/testing/selftests/powerpc/utils.c b/tools/testing/selftests/powerpc/utils.c
index ed62f4153d3e..5c7dbd693584 100644
--- a/tools/testing/selftests/powerpc/utils.c
+++ b/tools/testing/selftests/powerpc/utils.c
@@ -127,6 +127,26 @@ bool is_ppc64le(void)
return strcmp(uts.machine, "ppc64le") == 0;
}
+int read_sysfs_file(char *fpath, char *result, size_t result_size)
+{
+ char path[PATH_MAX] = "/sys/";
+ int rc = -1, fd;
+
+ strncat(path, fpath, PATH_MAX - strlen(path) - 1);
+
+ if ((fd = open(path, O_RDONLY)) < 0)
+ return rc;
+
+ rc = read(fd, result, result_size);
+
+ close(fd);
+
+ if (rc < 0)
+ return rc;
+
+ return 0;
+}
+
int read_debugfs_file(char *debugfs_file, int *result)
{
int rc = -1, fd;
--
2.20.1
More information about the Linuxppc-dev
mailing list